Remote physiological signal recovery with efficient spatio-temporal modeling.

Area of Science:

• Biomedical Engineering
• Signal Processing
• Machine Learning

Background:

• Contactless physiological monitoring via remote photoplethysmography (rPPG) has significant applications in health and affective computing.
• Existing deep learning methods for rPPG are susceptible to motion and illumination artifacts, limiting their temporal characteristic exploitation.

Purpose of the Study:

• To develop an efficient spatiotemporal modeling-based method for robust rPPG signal recovery.
• To improve the accuracy and generalization of physiological measurements from video data.

Main Methods:

• Utilized 3D central difference convolution for temporal context modeling and Huber loss for robust intensity-level rPPG recovery.
• Employed a dual-branch structure with soft attention for motion and appearance modeling.
• Introduced a multi-task learning setting for joint cardiac and respiratory signal measurement.

Main Results:

• Achieved Pearson's correlation coefficient above 0.96 across three public datasets, outperforming state-of-the-art methods.
• Demonstrated strong generalization ability through cross-database and video compression experiments.
• Ablation studies confirmed the effectiveness and necessity of each proposed module.

Conclusions:

• The proposed spatiotemporal modeling approach significantly enhances rPPG-based physiological signal measurement accuracy and robustness.
• The method offers a promising solution for reliable contactless health monitoring and affective computing applications.